Structure-aware deep learning model for peptide toxicity prediction.

Area of Science:

• Computational biology
• Biotechnology
• Drug discovery

Background:

• Antimicrobial resistance (AMR) is a major global health threat, driving the need for novel therapeutics like antimicrobial peptides (AMPs).
• Traditional methods for assessing AMP toxicity are resource-intensive, hindering rapid development.
• Predictive models are needed to streamline the screening of potential AMP candidates.

Purpose of the Study:

• To develop and validate tAMPer, a multi-modal deep learning model for predicting peptide toxicity.
• To integrate sequence composition and 3D structural information for enhanced toxicity prediction.
• To reduce the time and cost associated with experimental toxicity screening of AMPs.

Main Methods:

• Developed tAMPer, a deep learning model utilizing graph neural networks and recurrent neural networks.
• Represented peptides using a graph-based approach, encoding ColabFold-predicted 3D structures.
• Extracted features from amino acid sequences and spatial interactions within peptide structures.
• Evaluated performance on a public protein toxicity dataset and a custom AMP hemolysis dataset.

Main Results:

• tAMPer achieved an F1-score of 68.7% on the AMP hemolysis dataset, surpassing the next best method by 23.4%.
• On the protein toxicity benchmark, tAMPer improved the F1-score by over 3.0% compared to existing state-of-the-art methods.
• The model effectively integrates sequence and structural data for accurate toxicity prediction.

Conclusions:

• tAMPer offers a significant advancement in predicting peptide toxicity, particularly for AMPs.
• This model can accelerate the discovery and development pipeline for new antimicrobial agents.
• tAMPer reduces the reliance on costly and time-consuming experimental toxicity assays.